Personalized information access via recommendation is a mainstay of today's online information systems. As online information sources become larger, more varied and more complex, new recommendation tools are needed that can integrate all of the dimensions of the data when making personalized recommendations. This project will develop new recommendation techniques appropriate to the scale and complexity of online social networks. In particular, it will study how different decompositions of a network can be used create multiple recommendation components, and how these components can be combined into a single weighted ensemble. The project will enable more effective and efficient generation of personalized recommendations, helping users find useful connections even in large and complex information networks.
This project will explore a recommendation approach known as the Weighted Hybrid of Low-Dimensionality Recommenders (WHyLDR), which has proved effective in recommendation for the social web. WHyLDR decomposes heterogeneous networks into collections of two-dimensional matrices on the basis of meta-path expansions, where a meta-path represents a composition of link relations over the network. The low-dimensional components are combined using learned weights. One challenge of this research is that the number of meta-paths is unbounded. Therefore, the project will explore analytic criteria by which to evaluate the relative value of components in a hybrid. The project will experiment with four complex heterogeneous network data sets drawn from very different areas: employment, travel, music and scientific bibliography, using a variety of recommendation tasks and evaluation metrics.
